Google Launches Gemini 4 Carbon, Focused on Cybersecurity and Long-Running Agents

Two weeks after Gemini 4 Argon comes Gemini 4 Carbon, which Google presents as its most powerful model. According to Google, it leads the Artificial Analysis index and achieves a 0% attack success rate in an indirect prompt injection test. It also says its internal agents have worked for days on end on multi-step tasks.

The rollout is staggered: first a group of cybersecurity defenders, then trusted partners, with prior testing by the US Government. Launch price: $2/$10 per million input/output tokens, which will rise to $4/$20.

Carbon ranks number 1 in version 5.0 of the recently updated Artificial Analysis Intelligence Index, with a score of 59. It shows significant gains in coding and agentic tasks, as well as in key capabilities for cybersecurity defense, achieving a firm lead on SEC-bench Pro (88.2%) for vulnerability detection, ExploitGym (54.8%) for exploitation and CWE-bench (71.0%) for patching. In internal testing, Carbon drives a considerable improvement in secure code generation, and in live coding environments with our own teams, it shows reduced rates of errors and faking.

Building on Gemini 4 Argon , launched two weeks ago, Carbon illustrates the breakneck pace at which we are pushing the frontiers of science. Our researchers and models work side by side to accelerate capability development while we strengthen our safety measures.

We are taking a meticulous approach to the launch and are actively participating in the US government’s voluntary process for pre-release model testing. Starting this week, we will make Carbon available to a subset of cybersecurity experts and gradually expand access to other trusted partners through our Fairwind program . We continue to gather feedback and refine safety measures before Carbon becomes available to developers, businesses and consumers through our product platforms as soon as possible.

Gemini 4 Carbon will launch at the same introductory price ¹ as Argon, at $2 per million input tokens and $10 per million output tokens, with cached input tokens at a 95% discount on the input token price.

Improving coding and reasoning for challenging long-horizon tasks.

Designed to tackle far more complex, long-horizon engineering problems, Carbon is built for sustained automated workflows. Internally, our teams have seen Carbon run for days to overcome obstacles and tackle complex, multi-stage engineering tasks, such as identifying the root cause and fixing subtle bugs in large codebases and low-level systems.

Based on user feedback, we focused on improving the model’s performance across a range of areas, such as coding, automated and scientific tasks, as well as full-stack web development and multimodal understanding. Thanks to improved quality of outputs and artifacts in web development, Carbon gained 89 Elo points and ranked #4 (1768) on WebDev Arena .

State-of-the-art cybersecurity performance to help build secure systems.

Gemini 4 Carbon marks a step change in cybersecurity capabilities. It achieves state-of-the-art performance across the entire cyber defense lifecycle: vulnerability detection, validation and remediation.

Our teams are already seeing major advances in cyber defense. In internal testing, Carbon achieves more than double the success rate in secure code generation compared with Argon. As another example, the Chrome team reported that Carbon performed more than 10 times better than another lab’s frontier model at generating high-quality patches. Wiz, our leading cloud security platform, also observes that Carbon leads in web penetration testing, a core capability for Wiz’s Scan for Good initiative.

Carbon excels on leading external benchmarks (evaluated without restrictions):

• Detecting real-world vulnerabilities: The key to cybersecurity lies in detecting real-world vulnerabilities. Carbon achieves an excellent 88.2% on SEC-bench Pro , autonomously uncovering complex bugs hidden in real software.

• Validating risk through exploitation: To prioritize which vulnerabilities to fix first, or whether to fix them at all, cyber defenders increasingly rely on models that verify which vulnerabilities are exploitable. Carbon sets a new record on the popular exploitation benchmark ExploitGym (54.8%).

• Root cause remediation: After identifying vulnerabilities, fixing the root cause is a fundamental step in the cybersecurity lifecycle. Carbon reaches 71.0% on Collinear’s CWE-bench platform , turning validated findings directly into working code fixes.

Advancing safety and security for frontier models

Gemini 4 Carbon builds on the same multi-layered safety architecture as Argon, which combines model training, real-time misuse defenses, prompt injection resistance and chain-of-thought monitoring. In particular, we made significant progress on model alignment with Argon, and in our Safety report we reported on its tendency to behave consistently . With Carbon, we set out to go a step further on alignment and trained the model to respond appropriately to complex or high-pressure tasks.

• Reducing shortcuts and exploiting loopholes: With Carbon, we further reduced shortcut attempts during training while ensuring high-quality coding performance. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, and cheating attempts fell to 0.4%, improving overall accuracy. In real coding environments with our internal teams, we observed that high-severity misalignment alerts dropped by almost 80%.
• Greater transparency and honesty: In our internal testing, Carbon roughly halves unsupported claims compared with Argon. It includes methodological caveats directly in the final output or holds blocked actions with a clear list of what is still required, instead of making up information or glossing over missing evidence.
• Better collaboration with the user amid ambiguity: instead of guessing or forcing a path forward, Carbon retrieves relevant sources and surfaces discrepancies to the user before acting; as a result, users should see a much better collaborative experience with Carbon.
• Delivering on promises without overreaching: Carbon simultaneously reduced overcommitment by 26% and underdelivery by 29%, resulting in much more reliable end-to-end execution of multi-stage plans, respecting project constraints and staying strictly within scope.
• Prompt injection resistance: While Argon previously led Gray Swan’s indirect prompt injection (IPI) benchmark in prompt injection resistance with a very low attack success rate of 0.7%, Carbon achieves 0.0%, withstanding every attack attempt evaluated.

Coming soon to developers, businesses and consumers.

Moving the industry forward requires acting quickly and, at the same time, with meticulous planning. With Gemini 4 Carbon, our goal is to make significant advances in software engineering and cybersecurity without compromising safety or reliability. We look forward to seeing how the first cybersecurity professionals use this model and hope to bring Carbon to developers, businesses and consumers soon.